Research On Loop Closure Detection Of Visual SLAM Based On Reduced-dimensional Bag-of-Words Model

SLAM(simultaneous localization and mapping),which means simultaneous localization as well as mapping,focuses on how to use the sensor information to confirm its own trajectory and create an accurate map of the environment when a mobile robot is operating in an unknown environment.This paper focuses on the problem of closed-loop detection of SLAM in depth.Due to the large range and long time movements of the robot which tend to cause accumulation of map errors,closed loop detection is required in order to validate the map and correct the information to reduce the accumulated errors.Some closed-loop detection algorithms based on bag-of-words model,the features used for training words are artificially set,which will reduce the accuracy of closed-loop,while the use of k-means clustering method requires the number of clusters to be specified in advance,which also reduces the characterization ability of words,and cannot meet the requirements of real-time and accuracy of closed-loop detection,so this paper proposes a closed-loop detection method with reduced-dimensional bag-of-words model.Related work is as follows:(1)Improved feature extraction and feature clustering methods in bag-of-words models.Environmental features are extracted using the lightweight network MixNet and a progressive feature clustering method is proposed to violently match the extracted adjacent image descriptors and divide the matched images into existing nodes and those that do not match become new nodes,then the top 300 descriptors are retained and the entire vocabulary is constructed using a bag-of-words framework,sorted by the scores of key points.The progressive clustering method overcomes the traditional bag-of-words model which requires a pre-specified number of clusters due to the use of the k-means algorithm,and improves word representation.Experimental results of closed-loop detection on the City Centre and New College datasets show that the proposed algorithm can detect closed loops well compared to other closed-loop detection algorithms,and still has good accuracy at higher recall rates,while real-time performance is improved.(2)The high dimensionality of the feature matrix extracted by the MixNet network leads to some invalid words,for which a feature dimensionality reduction process is proposed.It uses Kernel Principal Component Analysis(Kernel PCA)to map the input feature matrix to a high-dimensional space via a kernel function and then performs PCA to reduce the dimensionality and improve the representational power of the data.Comparative experiments with other closed-loop detection algorithms show that the proposed method improves the feature matching time by 35.9%.This paper addresses the traditional bag-of-words model feature extraction and clustering problem by using MixNet to improve it and reduce the dimensionality of the feature matrix by kernel principal component analysis for the problem of high dimensionality.The proposed algorithm was studied comparatively on a relevant dataset and the results show that fast and accurate closed-loop detection can be achieved.